Economic models of alignment and extrapolated volition

Abhimanyu Pallavi Sudhir, Long Tran-Thanh

Keywords: alignment, extrapolated volition, economics, mechanism design, information economics

Abstract: We propose a simple formal definition of AI alignment motivated by an economic analogy. Market failures (such as imperfect information and imperfect assurance) naturally correspond to misalignment modes: in particular, though we leave this for future work, we believe that a ``recursive information markets'' mechanism naturally leads us to a formalization of extrapolated volition in terms of a value-of-information expression.

Serve As Reviewer: ~Abhimanyu_Pallavi_Sudhir1

Confirmation: I confirm that I and my co-authors have read the policies are releasing our work under a CC-BY 4.0 license.

Submission Number: 23